Grok’s official SDK guide schedules default text watermarking for supported models from 7 November 2026. This is a readiness check for the announced change, not verification that it has already reached production.
The output documentation, updated 31 October, gives no activation hour or timezone. Its coverage includes global and regional text-generation endpoints, Responses and Chat Completions, structured outputs and tool arguments.
Record the response, not an assumption
The SDK guide says x-watermark: true identifies watermarked output; absent or other values indicate otherwise. Its response and stream objects expose a watermarked property. Check your client version and whether a gateway preserves the original response headers.
Our suggested acceptance test is a small set of non-sensitive requests: ordinary prose, a schema-constrained object and a tool call whose arguments are validated without executing an action. Record model ID, endpoint, request time and returned watermark status. Compare the parsed results with your application’s existing acceptance criteria.
Keep detection separate from authorship
The output documentation describes statistical detection that can return uncertain; longer text strengthens the signal. Detector access requires application.
That is not a basis for treating a pasted paragraph as proof of who wrote it. In your own application, preserve appropriate request records and distinguish an API response’s generation status from a later judgment about a copied or edited passage. Avoid using a single detector result to accuse a person.
Review the team decision before changing it
Admin controls apply per model across all team keys; some models may be locked on. Opt-out requires acknowledging a legal exception and responsibility for disclosure obligations. Changes can take five minutes.
Map the applications using that team before proposing a setting change. Name the person responsible for reviewing any acknowledgment and the person verifying the resulting behavior. A technical setting does not determine whether a particular use meets applicable legal requirements.
These are proposed checks, not tests we performed. We have not called the API, changed a team setting, accepted terms or evaluated detector accuracy. Check the linked coverage list for your exact model; this report concerns API text output, not a claim about every Grok product or media type.